Proceedings of the 10th Mediterranean Conference on Control and Automation - MED2002 Lisbon, Portugal, July 9-12, 2002.

OPTIMISATION METHODS FOR IMPROVING FED-BATCH CULTIVATION OF E. COLI PRODUCING RECOMBINANT PROTEINS
I. Rocha, E.C. Ferreira
Centro de Engenharia Biolgica, Universidade do Minho 4710057 Braga PORTUGAL e-mail: {irocha, ecferreira}@deb.uminho.pt

Keywords: Optimisation, fed-batch fermentation, E. 1 Introduction coli, gradient methods, genetic algorithms. Fermentation processes are commonly operated in fedbatch mode in order to prevent the accumulation of Abstract toxic substrates or products [10,23] thus allowing the achievement of higher product concentrations. The Two optimisation techniques for the fed-batch bacteria Escherichia coli is usually grown under this cultivation of high cell density Escherichia coli kind of operation due to the well-known negative effect producing recombinant proteins were compared. of acetate, which is produced when the substrate, An unstructured m odel for the growth, based on the glucose, is presented above certain concentrations General State Space Dynamical Model [1] was used to [7,22]. represent the four major metabolic pathways: oxidative E. coli has been intensively studied in the last decades growth on glucose, fermentative growth on glucose, and has been widely used for the production of oxidative growth on acetate, and maintenance. The biopharmaceuticals [5,9]. In this kind of processes, the dilution rate (dependent on the substrate feed rate) was importance of operating near an optimal solution is chosen as the input variable. evident due to the high value of the metabolites Recombinant protein production is known to be produced. proportional, in our system, to the biomass During a common fed-batch E. coli fermentation concentration. Thus, biomass productivity was chosen process, the system states change considerably, from a as the criterion to be maximized. low initial cellular concentration to very high biomass The two methods compared were a first order gradient and product concentrations. This dynamic behaviour method based on Pontryagins minimum principle and a motivates the development of optimisation methods to stochastic method based on the biological principle of find the optimal input trajectories during the feeding natural evolution, using a genetic algorithm. The former stage in order to improve the process. The typical input method revealed less efficient concerning to the variable in these processes is the substrate inflow rate computed maximum, and dependence on good initial as a function of time. values. One way to evaluate the process performance is the measurement of the final product concentration. For the production of recombinant proteins in E. coli this is

usually related with the final biomass concentration The growth of E. coli may be interpreted through four obtained. Thus, a simple objective, although not the main reactions [3,6]: most accurate, is the maximisation of the final biomass Respiratory growth on glucose: concentration. (3) Many publications focus on this kind of optimisation problem in other bioprocesses [4,12-14,18,19,23], Fermentative growth on glucose: while only few [9,11] describe the application of 2 optimisation algorithms in E. coli fermentation. k2 S X + k8C + k3 A (4) In this work two approaches were considered, namely the application of a gradient method and a genetic Respiratory growth on acetate: 3 algorithm. k 4 A + k6O X + k9 C (5)
k1S + k5O X + k 7 C
1

1 Process Modelling
Process simulation was conducted with a developed model derived from the general state space dynamical model described by Bastin and Dochain [1]. Generally, this approach can be described as representing the dynamics of n components and m reactions bioprocess:

Maintenance:
S + k m1 A + k m2O k m3C
4

(6)

where S, O, X, C, A represent sugar (glucose), oxygen, biomass, carbon dioxide, and acetate respectively (in the sequel S, O, X, C, and A mean concentrations); 1, 2, 3, and 4 are the growth rates; k i are the yield (stoichiometric) coefficients. The application of the general state space dynamical model to these reaction network leads to the following model:
X 1 k S 1 d A = 0 dt O k5 C k7 0 X 0 0 (7) 1 1 2 S DSin 0 -k m1 X D A + 0 0 - km 2 3 O OTR 0 4 km 3 C 0 CTR

d = K ( , t ) D + F Q dt

(1)

where Rn is a vector representing the state components; K Rnm is the yield coefficient matrix;
Rm is the growth rates vector; the vectors F and Q

1 k2 k3 0 k8

1 0 k4 k6 k
9

are the feed rates and the gaseous outflow rates, where CTR is the carbon dioxide transfer rate from respectively (F, Q Rn). The scalar D is the dilution liquid to gas phase, OTR is the oxygen transfer rate rate (D-1 is the residence time), which will be the from gas to liquid phase, and Sin is the influent glucose manipulated variable, defined as follows: concentration. Depending on the feeding rate, E. coli growth can be taken simultaneously: i) from eq. 3 and eq. 6, ii) from W (t ) eqs. 3, 4, and 6 or iii) from eqs. 3, 5, and 6. For the maximum biomass production, the preferred metabolic -1 being Fin the influent flow rate (kg h ), and W the pathways correspond to the first case. Besides the lower biomass yield, the fermentative pathway (eq. 4) culture medium weight (kg). causes the production of acetate, an inhibitor of
D(t) = Fin ( t )

(2)

biomass growth at relatively small concentrations (7.5 g/kg) [7,22]. Acetate consumption (eq. 5) yields a smaller biomass yield than glucose oxidative consumption (eq. 3).

objective function to be a function of the states in every instant, which is not the case.

For both optimisation methods some parameters and constraints were based on a real E. coli fermentation In order to prevent the fermentation of glucose, it is system described elsewhere [16,17]. The final time necessary to keep the feeding at small values. On the was considered to be 25 hours, and the maximum feed other hand, those values must be sufficiently high in rate was 0.4 kg/h. order to achieve the desired biomass concentration in a short time. In the model, this behavior is governed by process kinetics, that is to say, from equations describing 1, Gradient methods [2] based on Pontyagrins minimum 2, 3, and 4. These equations and parameters k i used principle have been applied to the optimisation of generic bioprocesses [20] or hybridoma cells [18]. It is were based on the work described by [6]. reported that, although a good final result are normally achieved, computation effort is considerable. Another 3. Optimisation aspect that has to be taken in consideration is the initial guess of the optimisation problem. According to [8], when the objective function is linear, which is the Independently of the microorganism growth, the present case, the trajectories are v ery sensitive to the optimisation problem of a fed-batch fermentation is a initial conditions. This author suggests, as a solution to typical problem of dynamic optimisation, consisting on this disadvantage, the definition of a quadratic cost finding a sequence of values for a control variable that criterion. minimize a chosen objective function, subject to the For the implementation of this method, the MATLAB constraints represented by the dynamical model. 6.0 function fmincon was used. This function Thus, in this particular case, the problem consists on calculates the minimum of a function subject to several finding a sequence of values the vector D that kinds of constraints starting at an initial estimate. It uses maximize the objective function: Sequential Programming methods, in which a Quadratic Programming (QP) subproblem is solved at each J ( t f ) = X (t f ) (8) iteration [21]. An estimate of the Hessian of the Lagrangian is updated at each iteration using the BFGS subject to the constraints represented on eq. 7. formula. It requires as inputs the objective function, the Several optimisation methods have been applied to vector of initial estimates, constraints, upper and lower solve this kind of problem among them the two bounds for the variables and a previously defined methods proposed in this paper, a first order gradient vector of options. method based on Pontryagins minimum principle and a genetic algorithm. Another popular method is the so- The objective function was evaluated in a script .m called Dynamic Programming. Although it is still not file, a routine that calculated the final biomass clear if this method has some advantages over other concentration as a function of the input vector D in an th optimisation method in the particular case of fed-batch iterative way using 4 order Runge-Kutta to solve the fermentations [18], it has been applied to the differential equations of the model represented in eq. monoclonal antibody production optimisation [4] and (7). to Zymomous mobilis fermentation [23]. This method In order to optimise the computation time, the step of was not applied to this particular case, as it requires the the integration was different from the length of the input D. Thus, vector D was divided in 25 elements (1 h-1)

3.1 Gradient Method

that were interpolated 10 times for the calculation of the objective function value. There were no constraint functions passed to the function fmincon, as the final weight restriction (5 kg) was accounted by a penalty in the objective function. The upper and lower bounds of the variable D were zero and the maximum feed rate, respectively.

best. Variation in the population is achieved by genetic cross-over and mutation.

These algorithms proved to be very suitable for the optimisation of highly non-linear problems with m any variables. The application of these methods to fedbatch fermentations was done by several authors. Roubos et al. [19] and Nguang et al. [14] applied this The MATLAB routines were implemented in a Pentium approach to hybridoma cells, while yeast fermentation III computer running at 933 MHz in a Windows 2000 was the bioprocess studied by Na et al. [13]. environment. Several simulations were conducted with the routines described. The results are presented in Table 1. The function converged to a solution, although it took 100 iterations to achieve a high final biomass concentration. It should also be mentioned that the values shown were obtained for an initial vector D linearly increasing with time from 0 to 0.1. With an initial constant vector D (for example of 0.05) the function did not converge to a high objective function, remaining in the values of 5 g/kg even after 1000 iterations. This is due to the already mentioned sensitivity of the method to initial values with linear objective functions. After optimisation, the vector D corresponding to 1000 iterations was used to simulate a fermentation process during a 25 hours run. The results obtained are shown in figure 1.
60 50
X D

0.1 0.08 0.06 0.04 0.02 0

X (g/kg)

30 20 10 0 0 5 10 15 20 25

Time (h)

Figure 1. Evolution of the biomass and the input variable for the final solution of the gradient method (iteration 1000).

For the implementation of this method, a Genetic Algorithm Toolbox developed for MATLAB at the Number of Objective University of Sheffield was used [15]. It works with iterations function (g/kg) several genetic operators and supports binary, integer Initial values 7.61 and real-valued representations. The representation chosen for this work was the real-valued, due to the 100 51.03 advantages known [19]. The routine used as objective 1000 53.91 function was the same as in the previous method. The function tbxmpga was called from a high level script Table 1. Results obtained with the gradient and conducted all the operations, namely, initialisation method after 100 and 1000 iterations. of the population, ranking and selection based on the objective function values, recombination/crossover, mutation, evaluation, reinsertion and migration. The 3.2 Genetic Algorithm number of individuals evaluated in any iteration was Genetic Algorithms (GAs) have become a very 200. Each individual was a sequence of values Dk, with attractive method of non-linear optimisation in recent values between upper and lower bounds mentioned years. GAs are based on natural evolutionary above, composing a specific vector D. processes, where selection results in species that fit the

D (1/h)

40

The MATLAB routines were also ran in a Pentium III 4. Conclusions computer running at 933 MHz in a Windows 2000 The two methods under comparison revealed useful in environment. achieving an optimal solution for the fed-batch The results obtained are illustrated in Table 2. It is clear fermentation of E. coli. However, the GAs revealed that it took also 100 iterations to achieve a final solution more efficient concerning to the computed maximum, slightly higher than the one obtained after 1000 and dependence on good initial values. iterations with the gradient method. The simulation of The simple objective function studied in this work the fermentation with the vector D obtained gave the should be changed into a quadratic criterion in future results shown in figure 2. The shape of the input vector research in order to test the performance of the as well as of the biomass concentration profile along gradient method. the fermentation, are very similar to those observed for the previous method.

Acknowledgements
Number of iterations Initial values 50 100 Objective function (g/kg) 5.51 41.72 53.93
This work was supported by the PROTEXPRESS project (Agncia de Inovao). Fundao para a Cincia e a Tecnologia provided financial support for I. Rocha through a doctoral research grant (PRAXIS XXI/BD/16961/98).

Table 2. Results obtained with the GA method after 50 and 100 iterations.

References
[1] G. Bastin, D. Dochain. On-line Estimation and Adaptive Control of Bioreactors. Elsevier Science Publishers, Amsterdam, (1990).

60
X

0.1
D

50

0.08

X (g/kg)

40

D (1/h)

0.06 30 0.04 20 10 0 0 5 10 15 20 25 0.02 0

[2] A.E. Bryson, Jr. Dynamic Opimization, Addison-Wesley Longman, Inc., Menlo Park, 1999. [3] A. Cockshott, I. Bogle. Modelling the effect of glucose feeding on a recombinant E. coli fermentation, Bioprocess Engineering, 20, pp. 83-90, (1999). [4] M. de Tremblay, M. Perrier, C. Chavarie, J. Archambault, Fed-Batch Culture of Hybridoma Cells: Comparison of Optimal Control Approach and Closed Loop Strategies, Bioprocess Engineering, 9, pp. 13-21, (1993). [5] J. Fieschko, T. Ritch, Production of human alpha consensus interferon in recombinant Escherichia coli, Chemical Engineering Communications, 45, pp. 229-240, (1986).

Time (h)

Figure 2. Evolution of the biomass and the input variable for the final solution of the GA (iteration 100).

[6] V. Galvanauskas, R. Simutis, N. Volk, A. Lbert. [16] I. Rocha, E. C. Ferreira, An Integrated System Model Based Design of a Biochemical cultivation for Advanced Monitoring and Control of Fedprocess, Bioprocess Engineering, 18, pp. 227batch Fermentations of Recombinant E. coli, 8th 234, (1998). International Conference on Computer Applications in Biotechnology (D. Dochain, M. [7] K. Han, H. C. Lim and J. Hong, Acetic acid Perrier, Eds.), Qubec City, Canada, pp 353formation in Escherichia coli fermentation, 358, (2001). Biotechnology and Bioengineering, 39, pp. 663[17] I. Rocha, E. C. Ferreira, On-line simultaneous monitoring of glucose and acetate with flow[8] A. Johnson, Optimization of fed-batch injection analysis during high-cell-density biotechnical process, IFAC Modeling and fermentation of recombinant Escherichia coli, Control of Biotechnical Processes, Pergamon Analytica Chimica Acta, (accepted), (2002). Press, pp. 243-246, (1992). [9] S.Y. Lee, High cell-density culture of Escherichia [18] J.A. Roubos, C.D. de Gooijer, G. van Straten, A.J.B. van Boxtel, Comparison of optimization coli, Trends in Biotechnology, 14, pp. 98-105, methods for fed-batch cultures of hybridoma (1996). cells, Bioprocess Engineering, 17, pp. 99-102, [10] J. Lee, S.Y. Lee, S. Park, A.P.J. Middelberg (1997). Control of fed-batch fermentations, Biotechnology Advances, 17, pp. 29-48, (1999). [19] J.A. Roubos, G. van Straten, A.J.B. van Boxtel An evolutionary strategy for fed-batch bioreactor [11] J. Lee, W.F. Ramirez, Optimal fed-batch control optimization; concepts and performance, Journal of induced foreign protein production by of Biotechnology, 67, 173-187, (1999). recombinant bacteria A.I.Ch.E. J. 40, pp. 899[20] S. Shioya, Optimization and control for maximum 907, (1994). production in fed-batch culture, IFAC Modeling [12] D. Letan, A. Podgornik, A. Perdih, R. Stravs, and Control of Biotechnical Processes, The sequential algorithm for dynamic optimization Pergamon Press, pp. 117-122, (1992). of bioprocesses, IFAC Modeling and Control of Biotechnical Processes, Pergamon Press, pp. [21] The Mathworks Inc. Optimization Toolbox Users Guide. The Mathworks Inc, Natick, MA, 271-274, (1992). (2000). [13] J.-G. Na, Y.K. Chang, B.H. Chung, H.C. Lim, Adaptive optimization of fed-batch culture of [22] M. van de Walle, J. Shiloach, Proposed mechanism of acetate accumulation in two yeast by using genetic algorithms, Bioprocess recombinant Escherichia coli strains during high and Biosystems Engineering, 24, pp. 299-308, density fermentation, Biotechnology and (2002). Bioengineering, 57, pp. 71-78, (1998). [14] S.K. Nguang, L.Z. Chen, X.D. Chen, Optimisation of fed-batch culture of hybridoma [23] F.S. Wang, C.H. Shyu, Optimal feed policy for fed-batch fermentation of ethanol production by cells using genetic algorithms, ISA Transations, Zymomous mobilis, Bioprocess Engineering, 40, pp. 381-389, (2001). 17, pp. 63-68, (1997). [15] H. Pohlheim, GEATbx: Genetic and Evolutionary Algorithm Toolbox for use with MATLAB [24] T. Yamane, S. Shimizu, Fed-batch Techniques in Microbial Processes, Advances in Biochemical Documentation, (1995). www.systemtechnik.tuEngineering, 30, pp. 147-194, (1984). ilmenau.de/~pohlheim/GA_Toolbox/index.html 671, (1992).